Skip to content

[release/2.11] Skip known-failing ROCm 7.14 and MemPool tests; fix test_typing name_fn - #3599

Open
dnikolaev-amd wants to merge 3 commits into
release/2.11from
dnikolaev/rel2.11_fixes
Open

[release/2.11] Skip known-failing ROCm 7.14 and MemPool tests; fix test_typing name_fn#3599
dnikolaev-amd wants to merge 3 commits into
release/2.11from
dnikolaev/rel2.11_fixes

Conversation

@dnikolaev-amd

Copy link
Copy Markdown

Summary

Cherry-pick targeted CI stabilizations onto release/2.11 for TheRock / ROCm 7.14 testing. Test skips and test-infra helpers only.

  • MemPool + expandable segments (#185013): skip test_mempool_no_split and test_nested_mempool on Linux while MemPool segment counting is incompatible with expandable_segments (#180325, #180367).
  • ROCm 7.14 skips (#3525): add lazy_skip_if, skipIfRocmVersionAtLeast, and skip_if_rocm_ver_atleast_multiprocess; skip known-failing export, serialization, utils, and distributed tests on ROCm ≥ 7.14.
  • test_typing (#192648): fix parametrized name_fn to strip .py from subtest names (fixes too many values to unpack when running individual tests). Collapse requirements-ci.txt to a single mypy==1.16.0 pin so check_mypy_version.py works.

Test plan

  • python test/test_cuda.py TestMemPool.test_mempool_no_split TestMemPool.test_nested_mempool — skipped on Linux
  • PYTORCH_TEST_WITH_ROCM=1 python test/test_serialization.py TestSerialization.test_debug_set_in_ci — skipped on ROCm 7.14+
  • PYTORCH_TEST_WITH_ROCM=1 python test/test_utils.py TestStandaloneCPPJIT.test_load_standalone — skipped on ROCm 7.14+
  • PYTORCH_TEST_WITH_ROCM=1 python test/export/test_export_opinfo.py TestExportOnFakeCudaCUDA — skipped (9 tests)
  • PYTORCH_TEST_WITH_ROCM=1 python test/distributed/test_ce_colls.py NCCLCopyEngineCollectives.test_ce_allgather NCCLCopyEngineCollectives.test_ce_alltoall — skipped
  • PYTORCH_TEST_WITH_ROCM=1 python test/distributed/test_symmetric_memory.py SymmetricMemoryTest.test_subgroup SymmMemCollectiveTest.test_one_shot_all_reduce SymmMemCollectiveTest.test_two_shot_all_reduce — skipped
  • python test/test_typing.py TestTyping — 18/18 pass

…orch#185013)

Skip test_mempool_no_split and test_nested_mempool on Linux while MemPool
segment counting is incompatible with expandable_segments.

(cherry picked from commit 9661ae6)
Cherry-pick targeted hunks from 6296fc4 for TheRock CI failures:
add skipIfRocmVersionAtLeast / skip_if_rocm_ver_atleast_multiprocess guards
and skip export, serialization, utils, and distributed tests on ROCm 7.14+.

(cherry picked from commit 6296fc4)
…torch#192648)

Generated test names ending in ".py" break unittest name-based loading
(too many values to unpack). Use splitext in name_fn so subtests are
addressable by name.

Use a single mypy==1.16.0 pin in requirements-ci.txt so
check_mypy_version.py can resolve the CI mypy version.

(cherry picked from commit 51154e2)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant